Configuration of TensorFlow ObjecDetectionAPI under Anaconda3 of win10 system

Reference Blog 1: https://blog.csdn.net/zhaoyoulin2016/article/details/80615687

Reference Blog 2: https://blog.csdn.net/Zlase/article/details/78734138

Because it was installed with the Anaconda3 integrated package when python was installed before, the python in Anaconda3 is version 3.6, which is incompatible with tensorflow, so you need to create a python35 environment when using tensorflow. You can refer to Anaconda3 for building tensorflow.material.It is also for this reason that I made an error when adding a path when configuring some modules involved in the API, which will be described in detail below.

I. Installation of TensorFlow

(This step is omitted: you can refer to the blog: https://blog.csdn.net/r1254/article/details/76735740)

Second, download TensorFlowModels

Download link: https://github.com/tensorflow/models

After downloading, unzip it, and the research, samples, official and other folders will appear.Here, you need to add some modules to be used in the later test program to the path, and add the research in the decompressed package and the slim in the folder to the index path.

Method: It should be noted here that if you installed the python3.6 version with Anaconda3 and used TensorFlow to create the python35 environment, you need to go to the python35 folder and add the abovetwo paths.Create a new .pth file in Anaconda3>envs>python35 (the tensorflow environment I created, the name can be different)>Lib>site-packages, .pth fileThe name can be chosen at will, and the absolute paths of research and slim are written in it, as shown below:

Three, install protobuf

Download address: https://github.com/google/protobuf/releases

Some blogs mentioned that the downloaded version is the compressed package of protoc-3.4.0-win32.zip. There may be problems with other versions, and the specific reasons are not known.After decompression, copy protoc.exe in the bin folder to models-master\research. For the sake of safety, add the bin folder to the computer system environment variables (Create a new path in the path under the computer system variable and add ...\bin).

Then enter the computer cmd command window, switch to the TensorFlow environment (use the command activate python35), mine is activate python35.Then enter the research directory in the model-master from the national cd command, and execute the following command:

protoc object_detection/protos/*.proto --python_out=.

It is successful if no error is reported after a pause of one or two seconds. After success, some .PROTO files will appear in the models-master>research>object_detection>protos directory.

Then execute the following command in the research>object_detection>builders directory of the cmd command window:

python model_builder_test.py

If no error is reported, the following result will appear after waiting for more than ten seconds:

Indicates that the protoc configuration is successful, and the following code testing phase can be performed.

Four, test TensorFlowObjectDetectionAPI

First activate the TensorFlow environment through activate python35 in the cmd command window, then enter the models-master directory, and enter the command jupyter notebook to call up the web version python codeDebug the IDE, enter the research>object_detection directory and find object_detection_tutorial.ipynb, as followsFigure:

Click object_detection_tutorial.ipynb, the code debugging interface will appear, as shown below:

Then run Run All, wait for the running result, a UserWarning warning may appear in the middle, you can leave it alone, it will not affect the running result.The process will be affected by the network, and it may take a long time to get the result. You can observe the python3 in the upper right corner. It is a hollow circle when it is not running normally, and a solid circle when the program is running. Put the mouse on it to display the kernelbusy.The results are as follows:

Indicates that the test is successful. The above two pictures are the pictures that come with the number test program. You can replace it with your own pictures for testing. Here I put my own pictures into the program to run, and the results come out.as follows:

The above is for your reference. There may be flaws in the article. Please advise. In the next stage, I plan to use TensorFlow to train the targets in my own road scene. I am currently learning the data processing part of TensorFlow input images.